Benchmark Almanac

ProgramBench tests passed

Muse Wonder Full 88.5% leads since October 2026
Claude Opus 5 93% leads since July 2026
In use

Agents rebuild 200 programs from only their compiled binaries and docs. The score is the share of hidden behavioral tests the rebuilt program passes, averaged over tasks.

The board calls this an unofficial metric and ranks by fully resolved tasks, which the ProgramBench record holds. Labs report this number. Anthropic runs 166 of the 200 tasks.

Released
May 2026
Kind
Building whole projects
Level
Research frontier
Scored
Checked automatically
Answers
Code
Tasks
All public
Strongest model tested
Claude Fable 5.2 #1 on the capability index
Last in a launch post
September 2026
Last model added
November 2026
Scores refreshed
November 12th, 2026

In the authors’ words

Turning ideas into full software projects from scratch has become a popular use case for language models. Agents are being deployed to seed, maintain, and grow codebases over extended periods with minimal human oversight. Such settings require models to make high-level software architecture decisions. However, existing benchmarks measure focused, limited tasks such as fixing a single bug or developing a single, specified feature. We therefore introduce ProgramBench to measure the ability of software engineering agents to develop software holisitically. In ProgramBench, given only a program and its documentation, agents must architect and implement a codebase that matches the reference executable's behavior. End-to-end behavioral tests are generated via agent-driven fuzzing, enabling evaluation without prescribing implementation structure. Our 200 tasks range from compact CLI tools to widely used software such as FFmpeg, SQLite, and the PHP interpreter. We evaluate 9 LMs and find that none fully resolve any task, with the best model passing 95% of tests on only 3% of tasks. Models favor monolithic, single-file implementations that diverge sharply from human-written code.

Abstract of ProgramBench: Can Language Models Rebuild Programs From Scratch? (arXiv:2605.03546, CC0)
25
Models scored on its own leaderboard
2
Launch posts list it, from 2 labs
10
System cards report it, from 3 labs

Muse Wonder Full leads with 88.5%

Best score per model from the benchmark's own leaderboard, read on November 12th, 2026. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart.

Claude Opus 5 leads with 93%, claims included

Best score per model from every source, including labs’ own numbers and claims Epoch AI collected from reports and papers. Labs pick their own settings, so a claim can beat an independent run of the same model.

25 models. Scroll sideways for the rest.

  1. 88.5 Muse Wonder Full
  2. 87.7 Claude Opus 5.5
  3. 86.9 GPT-6 Astra
  4. 86.5 Muse Wonder
  5. 82.5 Claude Fable 5.1
  6. 74.7 Claude Opus 5
  7. 70.9 Claude Opus 4.8
  8. 70.8 Muse Spark 1.3
  9. 69.9 GPT-5.6 Sol
  10. 69.5 GPT-5.5
  11. 64.6 GLM-5.2
  12. 61.2 Gemini 3.7 Flash
  13. 57.2 Muse Spark 1.2
  14. 55.7 Gemini 3.6 Flash
  15. 55.1 Claude Opus 4.7
  16. 53.6 Gemini 3.5 Flash
  17. 52.1 Claude Opus 4.6
  18. 47.5 Claude Sonnet 4.6
  19. 47.0 Muse Spark 1.1
  20. 37.7 GPT-5.4
  21. 36.4 Gemini 3.1 Pro
  22. 31.8 Gemini 3 Flash
  23. 30.0 Claude Haiku 4.5
  24. 16.4 GPT-5.4 Mini
  25. 16.0 GPT-5 mini
  1. 1 Muse Wonder Full 88.5%
  2. 2 Claude Opus 5.5 xhigh 87.7%
  3. 3 GPT-6 Astra max 86.9%
  4. 4 Muse Wonder xhigh 86.5%
  5. 5 Claude Fable 5.1 xhigh 82.5%
  6. 6 Claude Opus 5 xhigh 74.7%
  7. 7 Claude Opus 4.8 xhigh 70.9%
  8. 8 Muse Spark 1.3 max 70.8%
  9. 9 GPT-5.6 Sol xhigh 69.9%
  10. 10 GPT-5.5 xhigh 69.5%
  11. 11 GLM-5.2 64.6%
  12. 12 Gemini 3.7 Flash 61.2%
  13. 13 Muse Spark 1.2 xhigh 57.2%
  14. 14 Gemini 3.6 Flash 55.7%
  15. 15 Claude Opus 4.7 xhigh 55.1%
  16. 16 Gemini 3.5 Flash 53.6%
  17. 17 Claude Opus 4.6 52.1%
  18. 18 Claude Sonnet 4.6 47.5%
  19. 19 Muse Spark 1.1 xhigh 47%
  20. 20 GPT-5.4 37.7%
  21. 21 Gemini 3.1 Pro 36.4%
  22. 22 Gemini 3 Flash 31.8%
  23. 23 Claude Haiku 4.5 30%
  24. 24 GPT-5.4 Mini 16.4%
  25. 25 GPT-5 mini 16%

Source: Official ProgramBench tests passed leaderboard

31 models. Scroll sideways for the rest.

  1. 93.0 Claude Opus 5
  2. 91.2 Claude Opus 5.5
  3. 88.5 Muse Wonder Full
  4. 87.6 Claude Fable 5.1
  5. 86.9 GPT-6 Astra
  6. 86.5 Muse Wonder
  7. 85.6 Claude Fable 5.2
  8. 82.0 Claude Haiku 5.5
  9. 80.5 Step 5 Preview
  10. 79.7 Claude Sonnet 5.5
  11. 79.0 Claude Opus 4.8
  12. 70.8 Muse Spark 1.3
  13. 69.9 GPT-5.6 Sol
  14. 69.5 GPT-5.5
  15. 64.6 GLM-5.2
  16. 61.2 Gemini 3.7 Flash
  17. 57.2 Grok 4.5
  18. 57.2 Muse Spark 1.2
  19. 55.7 Gemini 3.6 Flash
  20. 55.1 Claude Opus 4.7
  21. 53.6 Gemini 3.5 Flash
  22. 52.1 Claude Opus 4.6
  23. 50.3 Seed2.1 Pro
  24. 47.5 Claude Sonnet 4.6
  25. 47.0 Muse Spark 1.1
  26. 37.7 GPT-5.4
  27. 36.4 Gemini 3.1 Pro
  28. 31.8 Gemini 3 Flash
  29. 30.0 Claude Haiku 4.5
  30. 16.4 GPT-5.4 Mini
  31. 16.0 GPT-5 mini
  1. 1 Claude Opus 5 93%
  2. 2 Claude Opus 5.5 91.2%
  3. 3 Muse Wonder Full 88.5%
  4. 4 Claude Fable 5.1 87.6%
  5. 5 GPT-6 Astra max 86.9%
  6. 6 Muse Wonder xhigh 86.5%
  7. 7 Claude Fable 5.2 85.6%
  8. 8 Claude Haiku 5.5 82%
  9. 9 Step 5 Preview 80.5%
  10. 10 Claude Sonnet 5.5 79.7%
  11. 11 Claude Opus 4.8 79%
  12. 12 Muse Spark 1.3 max 70.8%
  13. 13 GPT-5.6 Sol xhigh 69.9%
  14. 14 GPT-5.5 xhigh 69.5%
  15. 15 GLM-5.2 64.6%
  16. 16 Gemini 3.7 Flash 61.2%
  17. 17 Grok 4.5 57.2%
  18. 18 Muse Spark 1.2 xhigh 57.2%
  19. 19 Gemini 3.6 Flash 55.7%
  20. 20 Claude Opus 4.7 xhigh 55.1%
  21. 21 Gemini 3.5 Flash 53.6%
  22. 22 Claude Opus 4.6 52.1%
  23. 23 Seed2.1 Pro 50.3%
  24. 24 Claude Sonnet 4.6 47.5%
  25. 25 Muse Spark 1.1 xhigh 47%
  26. 26 GPT-5.4 37.7%
  27. 27 Gemini 3.1 Pro 36.4%
  28. 28 Gemini 3 Flash 31.8%
  29. 29 Claude Haiku 4.5 30%
  30. 30 GPT-5.4 Mini 16.4%
  31. 31 GPT-5 mini 16%

Sources: Labs’ model cards and official ProgramBench tests passed leaderboard

#ModelScoreSettingDateSource
1 Muse Wonder Full 88.5% mini-SWE-agent Oct 2026 Official leaderboard
2 Claude Opus 5.5 87.7% xhigh, mini-SWE-agent Sep 2026 Official leaderboard
3 GPT-6 Astra 86.9% max, mini-SWE-agent Sep 2026 Official leaderboard
4 Muse Wonder 86.5% xhigh, mini-SWE-agent Oct 2026 Official leaderboard
5 Claude Fable 5.1 82.5% xhigh, mini-SWE-agent Sep 2026 Official leaderboard
6 Claude Opus 5 74.7% xhigh, mini-SWE-agent Jul 2026 Official leaderboard
7 Claude Opus 4.8 70.9% xhigh, mini-SWE-agent May 2026 Official leaderboard
8 Muse Spark 1.3 70.8% max, mini-SWE-agent Sep 2026 Official leaderboard
9 GPT-5.6 Sol 69.9% xhigh, mini-SWE-agent Jul 2026 Official leaderboard
10 GPT-5.5 69.5% xhigh, mini-SWE-agent Apr 2026 Official leaderboard
11 GLM-5.2 64.6% mini-SWE-agent Jun 2026 Official leaderboard
12 Gemini 3.7 Flash 61.2% mini-SWE-agent Aug 2026 Official leaderboard
13 Muse Spark 1.2 57.2% xhigh, mini-SWE-agent Aug 2026 Official leaderboard
14 Gemini 3.6 Flash 55.7% mini-SWE-agent Jul 2026 Official leaderboard
15 Claude Opus 4.7 55.1% xhigh, mini-SWE-agent Apr 2026 Official leaderboard
16 Gemini 3.5 Flash 53.6% mini-SWE-agent May 2026 Official leaderboard
17 Claude Opus 4.6 52.1% mini-SWE-agent Feb 2026 Official leaderboard
18 Claude Sonnet 4.6 47.5% mini-SWE-agent Feb 2026 Official leaderboard
19 Muse Spark 1.1 47% xhigh, mini-SWE-agent Jul 2026 Official leaderboard
20 GPT-5.4 37.7% mini-SWE-agent Mar 2026 Official leaderboard
21 Gemini 3.1 Pro 36.4% mini-SWE-agent Feb 2026 Official leaderboard
22 Gemini 3 Flash 31.8% mini-SWE-agent Dec 2025 Official leaderboard
23 Claude Haiku 4.5 30% mini-SWE-agent Oct 2025 Official leaderboard
24 GPT-5.4 Mini 16.4% mini-SWE-agent Mar 2026 Official leaderboard
25 GPT-5 mini 16% mini-SWE-agent Aug 2025 Official leaderboard
#ModelScoreSettingDateSource
1 Claude Opus 5 93% Jul 2026 Lab model card
2 Claude Opus 5.5 91.2% Sep 2026 Lab model card
3 Muse Wonder Full 88.5% mini-SWE-agent Oct 2026 Official leaderboard
4 Claude Fable 5.1 87.6% Sep 2026 Lab model card
5 GPT-6 Astra 86.9% max, mini-SWE-agent Sep 2026 Official leaderboard
6 Muse Wonder 86.5% xhigh, mini-SWE-agent Oct 2026 Official leaderboard
7 Claude Fable 5.2 85.6% Nov 2026 Lab model card
8 Claude Haiku 5.5 82% Oct 2026 Lab model card
9 Step 5 Preview 80.5% Sep 2026 Lab model card
10 Claude Sonnet 5.5 79.7% Sep 2026 Lab model card
11 Claude Opus 4.8 79% May 2026 Lab model card
12 Muse Spark 1.3 70.8% max, mini-SWE-agent Sep 2026 Official leaderboard
13 GPT-5.6 Sol 69.9% xhigh, mini-SWE-agent Jul 2026 Official leaderboard
14 GPT-5.5 69.5% xhigh, mini-SWE-agent Apr 2026 Official leaderboard
15 GLM-5.2 64.6% mini-SWE-agent Jun 2026 Official leaderboard
16 Gemini 3.7 Flash 61.2% mini-SWE-agent Aug 2026 Official leaderboard
17 Grok 4.5 57.2% Jul 2026 Lab model card
18 Muse Spark 1.2 57.2% xhigh, mini-SWE-agent Aug 2026 Official leaderboard
19 Gemini 3.6 Flash 55.7% mini-SWE-agent Jul 2026 Official leaderboard
20 Claude Opus 4.7 55.1% xhigh, mini-SWE-agent Apr 2026 Official leaderboard
21 Gemini 3.5 Flash 53.6% mini-SWE-agent May 2026 Official leaderboard
22 Claude Opus 4.6 52.1% mini-SWE-agent Feb 2026 Official leaderboard
23 Seed2.1 Pro 50.3% Jun 2026 Lab model card
24 Claude Sonnet 4.6 47.5% mini-SWE-agent Feb 2026 Official leaderboard
25 Muse Spark 1.1 47% xhigh, mini-SWE-agent Jul 2026 Official leaderboard
26 GPT-5.4 37.7% mini-SWE-agent Mar 2026 Official leaderboard
27 Gemini 3.1 Pro 36.4% mini-SWE-agent Feb 2026 Official leaderboard
28 Gemini 3 Flash 31.8% mini-SWE-agent Dec 2025 Official leaderboard
29 Claude Haiku 4.5 30% mini-SWE-agent Oct 2025 Official leaderboard
30 GPT-5.4 Mini 16.4% mini-SWE-agent Mar 2026 Official leaderboard
31 GPT-5 mini 16% mini-SWE-agent Aug 2025 Official leaderboard

Labs report scores for 10 models in their cards. Switch the scores above to rank them together.

Of runs with a cost, the best scores 87.7% for $36 a task

Runs on the cost frontier score higher than any run that's cheaper. GPT-6 Astra at max comes within 0.8 points of the best for $11 a task, 3 times cheaper. The costs are the leaderboard's own estimates, so compare them only within this benchmark.

OpenAI Google DeepMind Meta Anthropic Cost frontier One model at each effort level
0% 20% 40% 60% 80% 100% $0.01 $0.10 $1.00 $10 $100 Cost per task, log scale Claude Opus 5.5, xhigh Muse Spark 1.3, max Muse Spark 1.3, xhigh GPT-5.6 Sol Muse Spark 1.1, xhigh GPT-5.4 Gemini 3 Flash 0% 20% 40% 60% 80% 100% $0.01 $0.10 $1.00 $10 $100 Cost per task, log scale Claude Opus 5.5, xhigh Muse Spark 1.3, max

Source: Official ProgramBench tests passed leaderboard

ModelEffortHarnessScoreCost per taskOutput tokens
Muse Wonder Full mini-SWE-agent 88.5% 352,279
Claude Opus 5.5 xhigh mini-SWE-agent 87.7% $36 434,511
GPT-6 Astra max mini-SWE-agent 86.9% $11 91,595
Muse Wonder xhigh mini-SWE-agent 86.5% $37 345,971
Claude Fable 5.1 xhigh mini-SWE-agent 82.5% $57 445,587
Claude Opus 5 xhigh mini-SWE-agent 74.7% $51 357,785
Claude Opus 4.8 xhigh mini-SWE-agent 70.9% $21 270,298
Muse Spark 1.3 max mini-SWE-agent 70.8% $6.46 387,723
GPT-5.6 Sol xhigh mini-SWE-agent 69.9% $6.08 56,179
GPT-5.5 xhigh mini-SWE-agent 69.5% $8.85 51,624
Muse Spark 1.3 xhigh mini-SWE-agent 68.6% $2.02 216,872
GPT-5.5 high mini-SWE-agent 66.5% $3.65 27,379
GLM-5.2 mini-SWE-agent 64.6% $25 237,624
Gemini 3.7 Flash mini-SWE-agent 61.2% $2.04 74,885
GPT-5.6 Sol mini-SWE-agent 57.8% $1.00 13,084
Muse Spark 1.2 xhigh mini-SWE-agent 57.2% $1.48 329,742
GPT-5.5 mini-SWE-agent 56.6% $1.21 11,928
Gemini 3.6 Flash mini-SWE-agent 55.7% $4.83 107,792
Claude Opus 4.7 xhigh mini-SWE-agent 55.1% $11 97,205
Gemini 3.5 Flash mini-SWE-agent 53.6% $5.60 132,679
Claude Opus 4.6 mini-SWE-agent 52.1% $11 89,147
Claude Opus 4.7 mini-SWE-agent 50.9% $3.81 41,669
Claude Sonnet 4.6 mini-SWE-agent 47.5% $27 190,227
Muse Spark 1.1 xhigh mini-SWE-agent 47% $0.73 357,115
GPT-5.4 mini-SWE-agent 37.7% $0.33 7,408
Gemini 3.1 Pro mini-SWE-agent 36.4% $1.51 18,071
Gemini 3 Flash mini-SWE-agent 31.8% $0.30 22,561
Claude Haiku 4.5 mini-SWE-agent 30% $0.80 47,506
GPT-5.4 Mini mini-SWE-agent 16.4% $0.04 3,147
GPT-5 mini mini-SWE-agent 16% $0.03 7,146

7 models have held number one on the official leaderboard since its release in May 2026 3 models have held number one with claims included since its release in May 2026

Lab claims count from the day the card came out, independent runs from the model’s release. We draw one row each for the benchmark’s own leaderboard and what labs claim in their cards.

OpenAI Anthropic Meta
Own leaderboard Passed 85% in Sep 2026, by GPT-6 Astra
Muse Wonder Full 88.5%
Lab claims Passed 85% in Jul 2026, by Claude Opus 5
Claude Opus 5 93%

Sources: Official ProgramBench tests passed leaderboard and labs’ model cards

Anthropic’s own number for Claude Opus 5 passed 85% in July 2026

All the scores we have, one dot each, so a model tested at several settings appears more than once.

On the benchmark’s own leaderboard Reported by the lab in its model card Best so far
0% 20% 40% 60% 80% 100% 2026 2027 Near the ceiling (85%) Claude Opus 5 Claude Opus 4.8 GPT-5.5 Claude Opus 4.7 Gemini 3 Flash GPT-5 mini 0% 20% 40% 60% 80% 100% 2026 2027 Near the ceiling (85%) Claude Opus 5 Claude Opus 4.8

Sources: Official ProgramBench tests passed leaderboard and labs’ model cards

2 labs put it in a launch post in the last six months

Launch posts from every lab that has listed it, one tick per post. Tall colored ticks include it, and short gray ones leave it out.

ByteDance Seed
StepFun
DateLaunch postScoreAs the post writes it
Sep 20, 2026 Step 5 Preview 80.5 ProgramBench (Pass Rate)
Jun 23, 2026 Seed2.1 Pro 50.3 ProgramBench (resolved / almost resolved / average pass rate) 0/1/50.3

10 system cards report it, from 3 labs

Besides launch posts, we read the system cards, model cards and technical reports labs publish with a release. They report far more, including safety and frontier-risk tests.

DateDocumentScoreAs the document writes itSection
Nov 9, 2026 Claude Fable 5.2 system card 85.6 ProgramBench hidden-test pass rate on 166 golden tasks, single agent, mini-swe-agent harness, no time limit, 1M-token window. Showing remaining context raises the score to 91.6% (Section 8.10.1) Capabilities
Oct 7, 2026 Claude Haiku 5.5 system card 82 ProgramBench, hidden test pass rate on 166 of 200 tasks, mini-swe-agent without time limit (8.7.1) Capabilities
Sep 28, 2026 Claude Sonnet 5.5 system card 79.7 ProgramBench hidden test pass rate on 166 tasks, single agent, mini-swe-agent harness (8.10.1) Capabilities
Sep 22, 2026 Claude Opus 5.5 system card 91.2 ProgramBench hidden test pass rate on 166 of 200 tasks, single agent, mini-swe-agent harness, no time limit (8.10.1) Capabilities
Sep 1, 2026 Claude Fable 5.1 system card 87.6 ProgramBench hidden test pass rate on 166 tasks, single agent, mini-swe-agent harness, no time limit (8.11.1) Capabilities
Jul 24, 2026 Claude Opus 5 system card 93 ProgramBench hidden test pass rate on 166 of 200 tasks after five 1M-token episodes, single agent (83% after the first episode) (8.9.1) Capabilities
Jul 14, 2026 Grok 4.5 model card 57.2 ProgramBench, tests-passed rate, high (2.8) Capabilities
Jun 30, 2026 Claude Sonnet 5 system card ProgramBench, hidden test pass rate on 166 of 200 tasks across 5 episodes, given only as a range 76 to 86%, no single value (8.8) Capabilities
Jun 23, 2026 Seed2.1 Pro model card 50.3 ProgramBench, resolved 0 / almost resolved 1 / average pass rate 50.3 (Table 5) Capabilities
May 28, 2026 Claude Opus 4.8 system card 79 ProgramBench hidden test pass rate on 166 of 200 tasks, one episode of up to 1M tokens, near-final snapshot. Text gives 79 to 88% across 1 to 5 episodes (8.5, Figure 8.5.A) Capabilities

The lab's number comes out ahead in 4 of 4 pairs that differ

Each score a lab reported for its own model, beside the same model's best score on the leaderboard. Gaps mostly come from different harnesses, effort levels and task subsets. Every benchmark.

Model Lab says Board says Gap What differs
Claude Opus 5 93% ProgramBench hidden test pass rate on 166 of 200 tasks after five 1M-token episodes, single agent (83% after the first episode) (8.9.1) 74.7% Leaderboard, xhigh, mini-SWE-agent +18 pts The lab used five million-token episodes on 166 filtered tasks, while the board evaluates mini-SWE-agent on the full benchmark.
Claude Opus 4.8 79% ProgramBench hidden test pass rate on 166 of 200 tasks, one episode of up to 1M tokens, near-final snapshot. Text gives 79 to 88% across 1 to 5 episodes (8.5, Figure 8.5.A) 70.9% Leaderboard, xhigh, mini-SWE-agent +8.1 pts
Claude Fable 5.1 87.6% ProgramBench hidden test pass rate on 166 tasks, single agent, mini-swe-agent harness, no time limit (8.11.1) 82.5% Leaderboard, xhigh, mini-SWE-agent +5.1 pts
Claude Opus 5.5 91.2% ProgramBench hidden test pass rate on 166 of 200 tasks, single agent, mini-swe-agent harness, no time limit (8.10.1) 87.7% Leaderboard, xhigh, mini-SWE-agent +3.5 pts

More in Coding

All 282 in Coding
  • SWE-bench Verified In use 500 human-validated GitHub issues from 12 Python repos. Agent writes a patch that must pass hidden tests. Claude Opus 4.7 83.5%
  • LiveCodeBench v6 In use Competitive programming problems collected after model cutoffs from LeetCode, AtCoder, Codeforces. Scored pass@1 on hidden tests. DeepSeek-V4-Pro 93.5%
  • SWE-bench Multilingual In use 300 GitHub-issue tasks across nine programming languages in SWE-bench format. Patch graded by repository tests. Claude Fable 5.2 96.6%
  • DeepSWE In use 113 original long-horizon engineering tasks across 91 repos and five languages, written from scratch. Hand-written verifiers accept alternative solutions. GPT-6 Astra 74.1%
  • SciCode In use Scientist-curated research coding problems split into subproblems across natural sciences. Code checked against test cases. Seed2.1 Pro 59.8%
  • Codeforces rating In use Model solves recent Codeforces contest problems. An Elo-style rating is computed from simulated contest performance. DeepSeek V4.1 Flash 3,471